Draft Room

• 

  Draft Room cover

Inspiration

• I have always been a big football and Fantasy Premier League fan, this is one of those challenge where I can have my agents play this with me seamlessly

What it does

• A fantasy football squad builder that is better with an agent in it: fourteen WebMCP tools, six of them with no button, for scenarios, formations, deterministic probability, and atomic plans.

How we built it

• Next.js 15 (static export), TypeScript, Tailwind CSS v4, shadcn/ui, Zustand, zod, Three.js WebGPU renderer via React Three Fiber, webmcp-types, Vitest. Fonts: Big Shoulders, Sofia Sans, Doto.

Challenges we ran into

• Lots of tools are being exposed but cannot instruct the agents to use them properly at the beginning

Accomplishments that we're proud of

• Say "compare a 4-3-3 built around Haaland with a 3-5-2 built around Saka; which is more likely to clear 1,800 points?" and the agent runs both plans through a seeded Monte Carlo with rotation risk and club correlation, reports the probabilities, and applies the winner in one atomic call. No UI could offer that as a form.

What we learned

• Most agent-on-a-page demos let the agent press the buttons a human already has. Draft Room gives the agent capability the page does not expose any other way: simulate_scenarios evaluates several hypothetical squads against a probability model without changing anything, apply_scenario commits one of them atomically, set_formation re-shapes the XI, project_points and player_outlook answer "how likely" rather than "how many", and find_upgrades ranks swaps under the remaining budget. Tools are registered with the native imperative API during HTML parse, carry readOnlyHint annotations that match what they do, and return a message sentence the agent can relay. Clicks and tool calls run the same store, so the page attributes every change to you (coral) or the agent (cyan), and a console logs every call with its input, result, and duration.

What's next for Draft Room

• We will develop this into a fully polished game with lots of feature parity and sports coverage, expect this to be a new kind of game where humans and agents can play together

Built With

• css
• next.js
• shadcn/ui
• tailwind
• three.js
• typescript
• webgpu
• zod
• zustand